drm/amdgpu: remove v_seq handling from the scheduler v2

Simply not used any more. Only keep 32bit atomic for fence sequence numbering.

v2: trivial rebase

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com> (v1)
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com> (v1)
Reviewed-by: Chunming Zhou <david1.zhou@amd.com> (v1)

drivers/gpu/drm/amd/amdgpu/amdgpu.h

@@ -1047,7 +1047,7 @@ int amdgpu_ctx_put(struct amdgpu_ctx *ctx);
 struct amdgpu_ctx *amdgpu_ctx_get_ref(struct amdgpu_ctx *ctx);
 
 uint64_t amdgpu_ctx_add_fence(struct amdgpu_ctx *ctx, struct amdgpu_ring *ring,
-			      struct fence *fence, uint64_t queued_seq);
+			      struct fence *fence);
 struct fence *amdgpu_ctx_get_fence(struct amdgpu_ctx *ctx,
 				   struct amdgpu_ring *ring, uint64_t seq);
 

drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c

@@ -866,11 +866,9 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
 			kfree(job);
 			goto out;
 		}
-		job->ibs[parser->num_ibs - 1].sequence =
+		cs->out.handle =
 			amdgpu_ctx_add_fence(job->ctx, ring,
-					     &job->base.s_fence->base,
-					     job->base.s_fence->v_seq);
-		cs->out.handle = job->base.s_fence->v_seq;
+					     &job->base.s_fence->base);
 		list_sort(NULL, &parser->validated, cmp_size_smaller_first);
 		ttm_eu_fence_buffer_objects(&parser->ticket,
 				&parser->validated,

drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c

@@ -236,17 +236,13 @@ int amdgpu_ctx_put(struct amdgpu_ctx *ctx)
 }
 
 uint64_t amdgpu_ctx_add_fence(struct amdgpu_ctx *ctx, struct amdgpu_ring *ring,
-			      struct fence *fence, uint64_t queued_seq)
+			      struct fence *fence)
 {
 	struct amdgpu_ctx_ring *cring = & ctx->rings[ring->idx];
-	uint64_t seq = 0;
+	uint64_t seq = cring->sequence;
 	unsigned idx = 0;
 	struct fence *other = NULL;
 
-	if (amdgpu_enable_scheduler)
-		seq = queued_seq;
-	else
-		seq = cring->sequence;
 	idx = seq % AMDGPU_CTX_MAX_CS_PENDING;
 	other = cring->fences[idx];
 	if (other) {
@@ -260,8 +256,7 @@ uint64_t amdgpu_ctx_add_fence(struct amdgpu_ctx *ctx, struct amdgpu_ring *ring,
 
 	spin_lock(&ctx->ring_lock);
 	cring->fences[idx] = fence;
-	if (!amdgpu_enable_scheduler)
-		cring->sequence++;
+	cring->sequence++;
 	spin_unlock(&ctx->ring_lock);
 
 	fence_put(other);
@@ -274,21 +269,16 @@ struct fence *amdgpu_ctx_get_fence(struct amdgpu_ctx *ctx,
 {
 	struct amdgpu_ctx_ring *cring = & ctx->rings[ring->idx];
 	struct fence *fence;
-	uint64_t queued_seq;
 
 	spin_lock(&ctx->ring_lock);
-	if (amdgpu_enable_scheduler)
-		queued_seq = amd_sched_next_queued_seq(&cring->entity);
-	else
-		queued_seq = cring->sequence;
 
-	if (seq >= queued_seq) {
+	if (seq >= cring->sequence) {
 		spin_unlock(&ctx->ring_lock);
 		return ERR_PTR(-EINVAL);
 	}
 
 
-	if (seq + AMDGPU_CTX_MAX_CS_PENDING < queued_seq) {
+	if (seq + AMDGPU_CTX_MAX_CS_PENDING < cring->sequence) {
 		spin_unlock(&ctx->ring_lock);
 		return NULL;
 	}

drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c

@@ -126,7 +126,6 @@ int amdgpu_ib_schedule(struct amdgpu_device *adev, unsigned num_ibs,
 	struct amdgpu_ring *ring;
 	struct amdgpu_ctx *ctx, *old_ctx;
 	struct amdgpu_vm *vm;
-	uint64_t sequence;
 	unsigned i;
 	int r = 0;
 
@@ -199,12 +198,9 @@ int amdgpu_ib_schedule(struct amdgpu_device *adev, unsigned num_ibs,
 		return r;
 	}
 
-	sequence = amdgpu_enable_scheduler ? ib->sequence : 0;
-
 	if (!amdgpu_enable_scheduler && ib->ctx)
 		ib->sequence = amdgpu_ctx_add_fence(ib->ctx, ring,
-						    &ib->fence->base,
-						    sequence);
+						    &ib->fence->base);
 
 	/* wrap the last IB with fence */
 	if (ib->user) {

drivers/gpu/drm/amd/amdgpu/amdgpu_sa.c

@@ -435,8 +435,8 @@ void amdgpu_sa_bo_dump_debug_info(struct amdgpu_sa_manager *sa_manager,
 				seq_printf(m, " protected by 0x%016llx on ring %d",
 					   a_fence->seq, a_fence->ring->idx);
 			if (s_fence)
-				seq_printf(m, " protected by 0x%016llx on ring %d",
-					   s_fence->v_seq,
+				seq_printf(m, " protected by 0x%016x on ring %d",
+					   s_fence->base.seqno,
 					   s_fence->entity->scheduler->ring_id);
 
 		}

drivers/gpu/drm/amd/amdgpu/amdgpu_sched.c

@@ -111,7 +111,6 @@ int amdgpu_sched_ib_submit_kernel_helper(struct amdgpu_device *adev,
 			kfree(job);
 			return r;
 		}
-		ibs[num_ibs - 1].sequence = job->base.s_fence->v_seq;
 		*f = fence_get(&job->base.s_fence->base);
 		mutex_unlock(&job->job_lock);
 	} else {

drivers/gpu/drm/amd/scheduler/gpu_scheduler.c

@@ -156,14 +156,12 @@ int amd_sched_entity_init(struct amd_gpu_scheduler *sched,
 			  struct amd_sched_rq *rq,
 			  uint32_t jobs)
 {
-	uint64_t seq_ring = 0;
 	char name[20];
 
 	if (!(sched && entity && rq))
 		return -EINVAL;
 
 	memset(entity, 0, sizeof(struct amd_sched_entity));
-	seq_ring = ((uint64_t)sched->ring_id) << 60;
 	spin_lock_init(&entity->lock);
 	entity->belongto_rq = rq;
 	entity->scheduler = sched;
@@ -179,8 +177,7 @@ int amd_sched_entity_init(struct amd_gpu_scheduler *sched,
 		return -EINVAL;
 
 	spin_lock_init(&entity->queue_lock);
-	atomic64_set(&entity->last_queued_v_seq, seq_ring);
-	atomic64_set(&entity->last_signaled_v_seq, seq_ring);
+	atomic_set(&entity->fence_seq, 0);
 
 	/* Add the entity to the run queue */
 	amd_sched_rq_add_entity(rq, entity);
@@ -299,8 +296,6 @@ static void amd_sched_process_job(struct fence *f, struct fence_cb *cb)
 	unsigned long flags;
 
 	sched = sched_job->sched;
-	atomic64_set(&sched_job->s_entity->last_signaled_v_seq,
-		     sched_job->s_fence->v_seq);
 	amd_sched_fence_signal(sched_job->s_fence);
 	spin_lock_irqsave(&sched->queue_lock, flags);
 	list_del(&sched_job->list);
@@ -421,15 +416,3 @@ int amd_sched_destroy(struct amd_gpu_scheduler *sched)
 	kfree(sched);
 	return  0;
 }
-
-/**
- * Get next queued sequence number
- *
- * @entity The context entity
- *
- * return the next queued sequence number
-*/
-uint64_t amd_sched_next_queued_seq(struct amd_sched_entity *c_entity)
-{
-	return atomic64_read(&c_entity->last_queued_v_seq) + 1;
-}

drivers/gpu/drm/amd/scheduler/gpu_scheduler.h

@@ -42,9 +42,7 @@ struct amd_sched_entity {
 	struct list_head		list;
 	struct amd_sched_rq		*belongto_rq;
 	spinlock_t			lock;
-	/* the virtual_seq is unique per context per ring */
-	atomic64_t			last_queued_v_seq;
-	atomic64_t			last_signaled_v_seq;
+	atomic_t			fence_seq;
 	/* the job_queue maintains the jobs submitted by clients */
 	struct kfifo                    job_queue;
 	spinlock_t			queue_lock;
@@ -72,7 +70,6 @@ struct amd_sched_fence {
 	struct fence                    base;
 	struct fence_cb                 cb;
 	struct amd_sched_entity	        *entity;
-	uint64_t			v_seq;
 	spinlock_t			lock;
 };
 
@@ -148,8 +145,6 @@ int amd_sched_entity_init(struct amd_gpu_scheduler *sched,
 int amd_sched_entity_fini(struct amd_gpu_scheduler *sched,
 			  struct amd_sched_entity *entity);
 
-uint64_t amd_sched_next_queued_seq(struct amd_sched_entity *c_entity);
-
 struct amd_sched_fence *amd_sched_fence_create(
 	struct amd_sched_entity *s_entity);
 void amd_sched_fence_signal(struct amd_sched_fence *fence);

drivers/gpu/drm/amd/scheduler/sched_fence.c

@@ -30,16 +30,19 @@
 struct amd_sched_fence *amd_sched_fence_create(struct amd_sched_entity *s_entity)
 {
 	struct amd_sched_fence *fence = NULL;
+	unsigned seq;
+
 	fence = kzalloc(sizeof(struct amd_sched_fence), GFP_KERNEL);
 	if (fence == NULL)
 		return NULL;
-	fence->v_seq = atomic64_inc_return(&s_entity->last_queued_v_seq);
+
 	fence->entity = s_entity;
 	spin_lock_init(&fence->lock);
-	fence_init(&fence->base, &amd_sched_fence_ops,
-		&fence->lock,
-		s_entity->fence_context,
-		fence->v_seq);
+
+	seq = atomic_inc_return(&s_entity->fence_seq);
+	fence_init(&fence->base, &amd_sched_fence_ops, &fence->lock,
+		   s_entity->fence_context, seq);
+
 	return fence;
 }